Satellite Mapping of 30 m Time-Series Forest Distribution in Hunan, China, Based on a 25-Year Multispectral Imagery and Environmental Features

Highlights

What are the main findings?

• A 25-year (1999–2023) forest mapping at 30 m resolution using multi-source Landsat series, DEM, and climate data.
• A deep learning framework integrates multi-temporal imagery and environmental factors for forest cover dynamics.

What is the implication of the main findings?

• Validation with 9000 manual samples and official statistics confirms high accuracy (OA > 92%) and reliability.
• Superior to existing products in capturing fine-scale spatial patterns and complex forest boundaries.

Abstract

Forests play a critical role in Earth’s ecosystem, yet monitoring their long-term, large-scale spatiotemporal dynamics remains a significant challenge. This study addresses this gap by developing an integrated framework to map annual forest distribution in Hunan, China, from 1999 to 2023 at a high resolution of 30 m. Our methodology combines multi-temporal satellite imagery (Landsat 5/7/8/9) with key environmental variables, including digital elevation models, temperature, and precipitation data. To efficiently reconstruct historical maps, training samples were automatically derived from a reliable 2023 forest product using a transferable logic, drastically reducing manual annotation effort. Comprehensive evaluations demonstrate the robustness of our approach: (1) Qualitative analyses reveal superior spatial detail and temporal consistency compared to existing global forest maps. (2) Rigorous quantitative validation based on ∼9000 reference samples confirms high and stable accuracy (∼92.4%) and recall (∼91.9%) over the 24-year period. (3) Furthermore, comparisons with government forestry statistics show strong agreement, validating the practical utility of the data. This work provides a valuable, accurate long-term dataset that forms a scientific basis for critical downstream applications such as ecological conservation planning, carbon stock assessment, and climate change research, thereby highlighting the transformative potential of multi-source data fusion and automated methods in advancing geospatial monitoring.

1. Introduction

Forests are vital components of the Earth’s ecosystem, playing critical roles in sustaining carbon cycles [1], regulating climate [2], and conserving biodiversity [3]. Globally, forests cover approximately 30% of the terrestrial land area, sequestering about 260 million tons of carbon each year and providing habitats for over 80% of terrestrial species [4]. The United Nations Food and Agriculture Organization (FAO) has highlighted the importance of forest protection and monitoring [5,6], a priority further underscored by an annual global forest loss of roughly 10 million hectares due to human activities [7]. Trends in forest cover, however, are far from uniform. Although the global forest area continues to decline overall, certain regions exhibit signs of partial recovery [8]. China has achieved a net increase in its total forest area. Over the past two decades, the increase in global greening area has been equivalent to the coverage of the Amazon Rainforest, with China making a prominent contribution. Although China accounts for only 6.6% of the world’s total vegetation area, it contributes 25% of the global net increase in vegetation leaf area. Meanwhile, China has achieved a continuous net growth in total forest area, emerging as the country with the largest and fastest greening progress globally. From 2015 to 2025, the annual net increase in China’s forest area reached 1.69 million hectares, accounting for approximately one-fourth of the world’s newly added greening area [9,10]. This net gain, however, masks severe natural forest loss in key biodiversity regions, where historical rates of deforestation were both rapid and extensive by global standards, jeopardizing local biodiversity and critical ecosystem services [11,12]. In contrast, temperate and boreal forests are gradually recovering, aided by reduced deforestation rates, afforestation initiatives, and natural regrowth [13,14]. Despite these regional gains, the overall global forest situation remains a pressing concern [15].

The inaccessibility of many forested regions, particularly in mountainous areas, presents significant challenges for ground-based surveys. Traditional field methods are widely recognized as time-consuming and labor-intensive, rendering them unsuitable for large-scale and temporally frequent monitoring requirements [16,17]. In response, remote-sensing technologies have matured into a powerful alternative over the past few decades. The utilization of satellite imagery from platforms such as Landsat and Sentinel-2 [17], enhanced by high-resolution data acquired from unmanned aerial vehicles (UAVs) [18], has gained widespread adoption. This technological evolution aligns with contemporary environmental policy frameworks, notably China’s “carbon peak and carbon neutrality” objectives [19] and the National Master Plan for Major Projects of Protection and Restoration of Key Ecosystems [20], all of which necessitate sophisticated monitoring capabilities. The integration of multi-source remote-sensing data has correspondingly expanded in forest and grassland monitoring applications. Spectral imagery offers a practical and economically viable approach for characterizing the distribution of tree species and forest types [21,22,23]. These developments have enabled the production of national- and global-scale map products at fine spatial resolutions (30 m and 10 m), facilitated by open-access Landsat and Sentinel data and advances in deep learning-based approaches. Previous studies have extensively investigated large-scale forest patterns and changes [24,25]. In parallel, the research community has developed a variety of open-access land-cover products to support such efforts [26,27,28], notable examples of which include: the China land-cover dataset (CLCD) [29], the global land-cover product with a fine classification system at 30 m (GLC_FCS30) [30], the global 30 m land cover dataset (GlobaLand30) [31], and the finer resolution observation and monitoring of global land cover (FROM-GLC2015) [32].

Early methodologies for processing remote-sensing imagery in forest monitoring predominantly employed pixel-based machine learning algorithms, including decision trees [33], support vector machines [34], and random forests (RF) [35]. These methods operate by leveraging spectral and statistical attributes of image pixels. The RF algorithm, for instance, enhances classification performance through ensemble learning from multiple decision trees [36]. Successful implementations abound, such as the integration of Landsat-based trend detection for disturbance and recovery (LandTrendr) with RF for detecting forest dynamics [37], and the application of RF with multi-feature inputs for forest-type mapping [38]. Comparative analyses confirm that these advanced tools enable robust forest monitoring, with methods like RF achieving performance levels comparable to artificial neural networks (ANN) [39]. Nevertheless, pixel-based machine learning approaches face a persistent constraint: their performance in complex scenarios is limited and critically dependent on manual feature engineering. Moreover, by taking individual pixels as the smallest learning unit, such algorithms cannot perceive contextual information from their neighborhoods, often resulting in fragmented mapping outputs. The advancement of deep learning-based methods has enabled the progressive application of convolutional neural networks (CNNs) to remote-sensing image analysis [40,41]. These models exhibit distinct advantages in large-area forest-type classification [42,43]. For instance, Three-Dimensional Convolutional Neural Networks (3D-CNNs) effectively harness rich spectral and spatial information from imagery, thereby enhancing classification accuracy. However, most deep learning frameworks are typically designed based on a single data source, such as high-resolution RGB imagery or multispectral data from an individual satellite, and the majority of studies rely on single-temporal data [44].

However, under the combined pressures of extreme climate, natural disasters, pests and diseases, urban expansion, and deforestation, forest resources in China exhibit distinct spatiotemporal dynamics in terms of quantity, quality, and spatial distribution. Such inherent complexity and uncertainty in their evolutionary processes not only hinder the refined management of forest resources but also highlight the urgency of targeted monitoring strategies. Accurate, large-scale, and long-term dynamic monitoring of these resources thus proves to be an indispensable technical strategy for addressing the aforementioned challenges.

Nevertheless, persistent shortcomings remain in existing research, particularly in elucidating the long-term, large-scale evolutionary patterns of forest resources in China and in developing optimized methodologies for multi-source data fusion monitoring [45]. Recent developments in forest mapping highlight multi-source data fusion as a prevailing trend. Integrating diverse data types—such as multi-spectral imagery from Landsat [46] and Sentinel-2 [47], radar data from Sentinel-1 and Advanced Land Observing Satellite (ALOS) [48], and LiDAR measurements [49]—effectively mitigates limitations of single-source data, including cloud obstruction, topographic shadows, and inadequate structural representation. Despite progress, two key challenges persist: (1) Limited integration of ecological or climate variables. Multisource auxiliary data—such as meteorological records (e.g., temperature, precipitation) and topographic parameters (elevation, slope) from Digital Elevation Models (DEMs) [41]—are essential for distinguishing species-specific habitats yet remain underexploited. (2) Trade-offs in spatial, spectral, and temporal resolution. Existing products often sacrifice either temporal continuity [50] or spatial-thematic detail [51], hindering dynamic analysis and fine-scale applications [52].

Based on a comprehensive understanding of the aforementioned challenges (e.g., inadequate integration of multi-source environmental data in long-term forest mapping, and labor-intensive manual annotation for historical time-series samples) and practical research needs (e.g., high-resolution forest datasets for subtropical carbon cycle modeling), this study focuses on mapping the long-term forest distribution (1999–2023) at 30 m spatial resolution in Hunan, China. As a representative subtropical region with high carbon sequestration potential, Hunan’s complex terrain and climatic variability introduce additional complexities to accurate forest monitoring. To address prevailing limitations in long-term forest monitoring (e.g., insufficient spatial detail of historical products and poor consistency with regional ecological attributes), the core objectives of this work are twofold: First, to develop a robust deep learning framework that systematically integrates multi-source heterogeneous data (including Landsat multispectral remote sensing imagery, topographic variables derived from digital elevation models (DEM), and gridded climatic datasets of temperature and precipitation) for accurate, long-term forest distribution mapping in subtropical mountainous regions, enabling the simultaneous capture of spatial patterns and temporal dynamics of forest cover. Second, to validate the reliability and applicability of the generated forest distribution product via a comprehensive multi-dimensional assessment framework, thereby providing a credible, high-resolution dataset to support regional carbon cycle modeling, ecological conservation planning, and forest resource management.

In general, this study makes three targeted contributions as follows to achieve these objectives mentioned above: (1) We propose a forest dynamic monitoring approach that embeds multi-source multi-modal data (Landsat 5/7/8/9 multispectral imagery, DEM-derived topographic features, temperature, and precipitation) into a unified end-to-end deep learning training pipeline. This design facilitates the concurrent extraction and fusion of remote sensing spectral features and environmental contextual information, thereby enhancing the model’s adaptability to the complex environmental conditions of subtropical mountainous areas. (2) We design a cross-temporal mapping workflow tailored to long-term forest cover inference: specifically, we leverage the well-validated contemporary GLC_FCS30 product (which exhibits high accuracy in recent years) to generate high-quality training samples for historical forest cover reconstruction. This approach not only circumvents the labor-intensive manual annotation bottleneck in long-term time-series analysis but also ensures the inter-annual consistency of sample labels. (3) We conduct a multi-dimensional validation of the final forest product using three complementary evaluation dimensions: (a) qualitative comparisons of spatial detail with mainstream global/regional forest map products; (b) quantitative point-level accuracy assessments based on a long-term manual annotation dataset (collected across diverse seasons and vegetation types); and (c) consistency verification against official forest resource survey reports from local ecological authorities. The results indicate that our product outperforms existing mainstream products in spatial detail, achieves point-level classification accuracy exceeding 90% for most vegetation types, and demonstrates strong alignment with official statistics on forest area dynamics, collectively affirming the scientific value and practical utility of the monitoring outcomes.

2. Materials and Methods

2.1. Study Area and Data

2.1.1. Study Area

We conducted a long-term forest mapping study in Hunan Province, China. The study area covers approximately 211,800 km² (Figure 1a,b) and is located in central-south China (108°47^′–114°15^′ E, 24°38^′–30°08^′ N), south of Dongting Lake and along the middle reaches of the Yangtze River. This region serves as a key ecological corridor connecting the Yangtze River Economic Belt and the Guangdong–Hong Kong–Macao Greater Bay Area, making it a critical area for regional ecological conservation and forest management. The region experiences a subtropical monsoon climate with four distinct seasons, characterized by abundant sunlight and heat. The mean annual temperature is 17.7 °C (Figure 1d), and annual precipitation ranges from 1200 to 1700 mm, concentrated between April and July (52.1% of the total; (Figure 1f). Topographically, the area exhibits a horseshoe-shaped pattern, with mountains on the east, south, and west, hilly landscapes in the center, and a lacustrine plain in the north. Elevation varies by approximately 2100 m (Figure 1e). These conditions support a diverse forest ecosystem, which displays vertical zonation with elevation. Planted forests—such as Chinese fir, masson pine, and bamboo—also constitute a significant proportion due to anthropogenic influence. In 2024, the forest coverage rate in Hunan reached 53.15% (https://lyj.hunan.gov.cn, accessed on 6 June 2025). The 2023 classification map of the study area is shown in (Figure 1c).

2.1.2. Landsat Images

The Landsat program, jointly managed by the National Aeronautics and Space Administration (NASA) and the United States Geological Survey (USGS), is one of the longest-running Earth observation initiatives, providing continuous global land-surface observations since 1972. We constructed a multi-temporal dataset using Landsat imagery (retrieved via the Google Earth Engine (GEE) platform) acquired at three-year intervals from 1999 to 2023. To ensure comprehensive temporal coverage and minimize seasonal biases in the subtropical monsoon climate of Hunan Province, imagery was acquired throughout the year. This strategy leverages optimal acquisition windows in winter, which feature drier conditions and significantly reduced cloud coverage, while also capturing phenological variations in other seasons. All scenes selected had cloud coverage below 30%. The dataset includes a substantial number of scenes across the study period, with counts of 125 (1999), 169 (2002), 190 (2005), 237 (2008), 168 (2011), 193 (2014), 183 (2017), 153 (2020), and 292 (2023). We utilized six spectral bands from each image: blue, green, red, near-infrared (NIR), and two shortwave infrared bands (SWIR1, SWIR2). Detailed spectral band specifications are provided in (

Table 1. Details of Landsat images employed in this study.

Satellites	Sensors	Time	Number of Image Frames	Spatial Resolution
Landsat 9	OLI-2	2021–2024	353	30 m
Landsat 8	OLI	2013–2024	413	30 m
Landsat 7	ETM+	1999–2023	2378	30 m
Landsat 5	TM	1999–2011	1113	30 m

) and the data volume per epoch is illustrated in (Figure 2).

2.1.3. Environment Factors

The distribution and composition of forest ecosystems are governed by a range of environmental factors.From a physical perspective, precipitation, air temperature, and DEMs do not act independently on forest growth. Instead, they collectively shape the distribution patterns, growth rates, and community structures of forests by regulating plant physiological metabolism, material-energy exchange, and habitat conditions. Integrating such variables into forest distribution frameworks can significantly enhance both the process efficiency and the accuracy of thematic outputs [53,54]. Previous studies [55,56] have demonstrated that incorporating topographic metrics—such as elevation, slope, and aspect—derived from DEMs can effectively resolve spectral confusion among forest types, with one study reporting classification accuracy up to 92.63% using Landsat time-series and topographic data [57]. In mountainous regions, topography strongly mediates local climate and soil conditions, thereby influencing species distribution and improving classification reliability when combined with satellite imagery [58]. Beyond terrain, climatic variables such as temperature and precipitation also play critical roles in determining forest structure and composition.

As the core input of forest ecosystem water cycles, precipitation regulates forest growth and community structure via two key pathways. First, precipitation infiltrates soil through canopy interception and understory penetration, is absorbed by roots, and is transported to leaves. It supports photosynthetic photolysis and maintains leaf turgor and stomatal opening for efficient CO₂ uptake; sufficient precipitation boosts photosynthesis and dry matter accumulation, while deficit induces stomatal closure, limiting photosynthesis or halting growth.

Second, its infiltration depth affects root distribution; as a solvent, it dissolves and transports N, P, and K to roots, and regulates soil temperature via evaporative cooling. Spatiotemporal variability drives species selection: humid regions (annual precipitation > 1000 mm) support complex communities; arid/semi-arid areas (<400 mm) only sustain drought-tolerant open woodlands/shrubs. As the core energy driver of forest growth, air temperature regulates enzyme activity and material-energy exchange by controlling molecular motion. The optimal photosynthetic temperature for most trees is 20–30 °C; temperatures below 5 °C or above 35 °C inhibit photosynthesis (via enzyme inactivation or denaturation). Respiration rises exponentially with temperature, and excessive respiratory consumption reduces net growth. Seasonal temperature shifts trigger phenological events (e.g., bud break at 10 °C, autumn leaf senescence), while accumulated temperature (daily means ≥ 10 °C) defines species latitudinal distribution boundaries.

These factors have been shown to differentially affect natural and planted forests, highlighting their utility in refining vegetation mapping efforts [59,60]. Environmental variables have been increasingly applied to forest extraction and other related research fields [61,62]. In this study, we incorporated multi-year (1999–2023) environmental variables, including DEM, mean annual temperature, and mean annual precipitation. The DEM used in this study was the 30 m resolution product from the Shuttle Radar Topography Mission (SRTM), accessed through the USGS EarthExplorer platform. Temperature and precipitation datasets, originally at 1 km spatial resolution, were acquired from the National Tibetan Plateau Data Center (TPDC) and resampled to 30 m using cubic convolution interpolation in ArcGIS.

2.2. Data Pre-Processing

All acquired Landsat imagery and environmental data underwent a systematic preprocessing workflow to generate high-quality datasets suitable for subsequent analysis.

Specifically, the Landsat imagery (covering Landsat 5/7/8/9) was retrieved and preprocessed via the Google Earth Engine (GEE) platform: We first filtered imagery by the study area boundary, temporal range (e.g., 1999–2023 at three-year intervals), and cloud coverage (≤30%) (consistent with the cloud mask operation using the quality assessment (QA) band in GEE). Then, we applied scale factor correction, cloud/shadow masking, and band renaming to the imagery, and merged multi-sensor Landsat collections (Landsat 4–9) to form a unified image collection. Finally, the merged imagery was composited (using the median method) and clipped to the study area extent. To mitigate atmospheric interference, clouds and cloud shadows were first masked for each scene using the quality assessment (QA) band. Subsequently, the acquired DEM, precipitation, and temperature datasets at 1-km resolution were resampled to 30 m to match the spatial resolution of the Landsat imagery. All remote-sensing imagery and resampled environmental data were then cropped (subset) to a common spatial extent and precisely aligned to the same coordinate system and geodetic datum. Finally, multiple feature variables—including spectral characteristics, DEM data, precipitation indices, and temperature—that shared identical geographic coordinates were integrated at the channel level (Figure 3a). This process consolidates multi-dimensional attributes within the information representation at each spatial location, thereby significantly enhancing the feature richness of individual pixels.

2.3. Time-Series Forest Mapping Framework

This section sequentially introduces the proposed framework and its training/inference process across long-time-series and large-scale domains, starting with an architectural overview (Figure 3b). As a semantic segmentation network with an encoder–decoder architecture, the U-Net is employed to process high-resolution remote-sensing imagery and to achieve automated provincial-scale forest distribution mapping. As shown in (Figure 3b), the model operates through three key components: encoder, decoder, and skip Connections. The encoder extracts multi-scale semantic features (e.g., forest texture and canopy structure) through successive convolution-activation-max pooling modules, producing downsampled feature maps. The decoder then restores spatial resolution via transposed convolution. Through skip connections, upsampled deep features are fused with corresponding shallow features from the encoder, effectively recovering fine spatial details (e.g., forest boundaries) lost during downsampling.

Finally, a 1 × 1 convolutional layer with softmax activation converts the refined features into class probability maps, enabling pixel-level forest distribution without manual feature engineering. During inference, the trained model automatically generates province-scale forest maps directly from input imagery.

Based on the preprocessed 2023 remote-sensing imagery and the GLC_FCS30 product, the scenes covering Hunan Province were first cropped according to administrative boundaries. The annually composited, fully covered imagery was then divided into 16 non-overlapping image patches (Figure 3b), each measuring 6000 × 6000 pixels. To meet the input requirements of the pre-trained network, each patch was further split into batch-ready tiles of 256 × 256 pixels, with an overlap of 128 pixels between adjacent tiles. These tiles were fed into the proposed U-Net deep learning model for training. By averaging probability values across overlapping regions and selecting the maximum probability for the forest category, adjacent predicted batches were seamlessly merged into the forest prediction results. The training process was optimized using the cross-entropy loss function. Finally, the merged extraction results were sequentially mosaicked to form a complete forest distribution map. Subsequently, the trained model was applied to perform pixel-wise inference on image data from 1999 to 2023 at three-year intervals. This process generated forest extraction results for historical years, thereby completing the forest mapping procedure.

2.4. Evaluation Setting

2.4.1. Qualitative Comparison Setting

(1)

Comparison with forest extraction results from four land cover products

We compare four 30 m forest products—CLCD [29], GLC_FCS30 [30], GlobeLand30 [31], and FROM_GLC2015) [32]—which employ diverse data and feature strategies (

Table 2. The mapping characteristics of several compared products in this study.

Product	Input Feature	Classifier	Time Series Coverage
Ours	DEM Precipitation Temperature Landsat-5,7,8,9	U-Net	1999–2023
CLCD	Elevation Slope and aspect Landsat-4,5,7,8	RF	1990–2019
GLC_FCS30	Landsat-4,5,7,8	RF	1985–2020
Globaland30	Landsat-5,7,8	Multi-Label Classifier	2000–2020
FROM GLC2015	Landsat-8	RF	2015

). The information on these comparison products is shown in (Table 2).

The CLCD employs a RF algorithm to produce annual land cover maps of China (1999–2019), which include nine distinct land cover classes. For the purpose of this comparative analysis, all non-forest categories were consolidated into a unified non-forest class. The GLC_FCS30 represents the first global 30 m resolution land cover product integrating continuous change detection, covering the period from 1985 to 2020. It implements a refined classification system comprising 35 categories, including ten detailed forest types differentiated by canopy density and species characteristics. In this study, these forest subtypes were aggregated into a single forest category to maintain comparability. GlobeLand30, another Chinese-developed global land cover product, offers 30 m resolution data spanning 2000 to 2020 with ten primary land cover classes. For our comparative framework, we reclassified these into binary categories: forest and non-forest. Similarly, the FROM_GLC2015 dataset, generated using RF classification, provides both first-level and second-level land cover categories. Our analysis utilized only its first-level forest classification, with all other land cover types designated as non-forest.

To validate the reliability and applicability of the Landsat-derived forest type classification for Hunan Province developed in this study, these four widely recognized land cover products were selected as reference datasets for systematic comparison (Figure 3c).

(2)

Comparison with the other three forest mapping products

In forest remote sensing mapping, benchmarking specialized forest products (beyond land cover datasets) is critical. This study references three key datasets. The 2021 Global 30 m natural/planted forest map (disturbance frequency sampling) [63], China’s 30 m forest stand height map (UAV LiDAR validation) [64], and the 2001–2020 Global 250-m forest management type map (carbon storage-focused) [65], supporting multi-scale forest research. Details regarding these comparative products are summarized in (

Table 3. Comparison of forest mapping products.

Product	Resolution	Data Source	Scope	Method	Category
Ours	30 m	Landsat-5,7,8,9	Hunan Province	Deep learning RF	Forest, Non-forest
Product 1	30 m	Landsat-4,8 Sentinel-1	Global	RF, Time-series change detection (CCDC)	Global plantation Natural forest mapping
Product 2	30 m	Sentinel-1 Landsat	China	Machine learning mixed-effects model LightGBM/XGBoost/CatBoost	Forest stand mean height mapping
Product 3	250 m	MOD13Q1 ALOS PALSAR	Global	RF, Change Detection (CCDC, SCBP)	Natural forest Plantation Oil Palm Plantation Agroforestry System

). The 2021 30 m global natural/planted forest product uses Landsat 4–8 SR and Sentinel-1 SAR data, integrating RF and CCDC for binary classification. Its 30 m resolution and global coverage enable forest inventory and planted forest expansion monitoring.

China’s 30 m stand height product fuses Sentinel-1 SAR and Landsat maximum NDVI data, applying MLME and gradient-boosting algorithms to retrieve arithmetic/weighted mean heights. Validated via UAV LiDAR tree segmentation, it improves the accuracy of large-scale stand parameters, aiding carbon stock estimation and growth assessment.

The 2001–2020 250-m global forest management product uses MOD13Q1 and ALOS PALSAR data, combining RF, CCDC, and SCBP to classify 6 forest types. It quantifies carbon dynamics across management types, supporting nature-based climate mitigation strategies.

2.4.2. Generating Long-Term Pixel-Level Validation Set

To verify the performance of deep learning methods in forest type distribution mapping, we collected sample points every three years from 1999 to 2023 in the study area and completed the preprocessing (Figure 3c). First, based on multi-temporal remote-sensing imagery of the study area, stratified random sampling was conducted, following the principle of spatial randomness while accounting for the distribution of major landform types. A total of 1000 sample points were set per year, resulting in a cumulative total of 9000 sample points over the 9-year period, as shown in (Figure 4). Parameter settings were specifically applied to ensure uniform spatial distribution of the sample points across different topographic conditions, thereby avoiding verification bias caused by local clustering or sparsity. Second, all 9000 random points were manually annotated via visual interpretation of Landsat imagery combined with auxiliary high-resolution imagery. Pixels within the vertical projection range of tree crowns were labeled as “forest,” while those corresponding to other land cover types—such as farmland, built-up areas, and water bodies—were categorized as “non-forest.”

This study also used special field survey data on forest, grassland, and wetland resources in Hunan Province collected in 2023, providing accurate ground-truth references for annotating remote-sensing sample points. As illustrated in (Figure 5), the core content and logic of the study are presented from two complementary perspectives: The technical workflow and regional tree species characteristics. The left part outlines the complete technical process from data acquisition to accuracy assessment. Remote-sensing images acquired at three-year intervals from 1999 to 2023 provide long-term time-series data for analyzing spatiotemporal forest changes, capturing both gradual transitions and abrupt disturbances in forest cover. Random point validation quantitatively evaluates the accuracy of forest identification at multiple spatial scales, ensuring the reliability of subsequent analyses and enabling robust temporal trend detection.

The right panel maps the distribution of dominant tree species across prefecture-level cities in Hunan Province. For instance, Cinnamomum camphora dominates in Changsha, while Acer palmatum is prevalent in Zhuzhou. The spatial variation of these species reflects a complex interplay between ecological adaptation and human management. Divergent biological traits—such as growth cycles, canopy structure, and environmental responses—underpin their distinct influences on forest spatiotemporal dynamics and spectral signatures. A prime example is Cunninghamia lanceolata, a fast-growing conifer with a short rotation cycle that rapidly forms closed-canopy forests under suitable conditions. In Huaihua City, where it is the primary commercial species, this trait drives relatively swift forest area expansion over short periods, creating distinct successional patterns in time-series data. In contrast, the slow-growing Ginkgo biloba maintains stable spectral features. In its dominant regions of Yongzhou and Shaoyang, forest area changes are more gradual, with significant shifts detectable only over decadal scales. This comparison demonstrates how species-specific traits modulate observable landscape dynamics. By integrating these species-specific growth characteristics with spatiotemporal dynamics through mechanistic modeling, we can better explain regional differences in the rates and magnitudes of forest change while accounting for management interventions.

This integrated approach substantially enhances the reliability of forest resource monitoring across heterogeneous landscapes, offering a more scientific and operational basis for forest inventory, forestry planning, and ecological conservation programs. Consequently, this methodology provides valuable insights for relevant authorities in formulating targeted management strategies that balance ecological integrity with socioeconomic needs. On the one hand, the field survey data provide a high-precision ground truth benchmark for sample points, covering multiple land cover types (e.g., forest, cropland, and construction land) and forest community attributes (including species composition, stand structure, canopy density, and growth status), which fundamentally ensures the accuracy and representativeness of random point annotations across diverse ecological zones. On the other hand, by coupling precise spatial locations and detailed field attributes with multi-temporal spectral and textural features from remote-sensing imagery, the field data support the construction of a comprehensive verification system that bridges scale gaps between plot measurements and pixel-level classifications. This integrated framework allows quantitative evaluation metrics—such as OA and recall—to not only be statistically reliable but also accurately capture real-world landscape heterogeneity across space and time, thereby strengthening the validity of model performance assessment under varying ecological conditions.

2.4.3. Statistical Data from the Local Government

To enhance the comprehensiveness and reliability of the deep learning model for forest distribution, this study incorporated official forest resource statistics as a key validation benchmark (Figure 3c). The dataset was acquired through Government Information Disclosure applications submitted to the Hunan Provincial Government and municipal forestry bureaus (e.g., Yiyang City). It contains detailed forest area information for all municipal-level administrative units within the study area, covering the period from 1999 to 2023—consistent with the timeline of sample point collection. The acquisition timeframe of this official data aligns with the three-year intervals of field sample surveys, ensuring temporal consistency between the two validation sources and providing a solid foundation for long-term trend analysis.

The forestry resource data were primarily drawn from the Bulletin on Land Greening in Hunan Province, an annual report issued by the Hunan Provincial Forestry Bureau. This bulletin systematically summarizes key aspects of the province’s land greening progress, forest resource dynamics, and ecological conservation achievements each year, and is widely regarded as a highly authoritative and reliable reference. Several factors underpin its credibility: First, as a provincial-level specialized agency, the Hunan Provincial Forestry Bureau maintains a systematic monitoring network, a professional technical team, and a well-established data collection system, all of which ensure the comprehensiveness and scientific rigor of the data. Second, in terms of methodology, the bulletin employs a combination of satellite remote sensing and field surveys to conduct multi-source monitoring and statistical analysis of forest resources across the province, thereby significantly improving data accuracy and objectivity. The integration of ground truthing with remote-sensing observations further enhances the reliability of the statistics. Third, the bulletin is disseminated through official channels such as the bureau’s website, guaranteeing standardized sourcing and high transparency. Moreover, its annual publication schedule is synchronized with the national Land Greening Bulletin of China, enabling effective cross-scale data comparison and policy alignment. The core indicator adopted from this official dataset is the total forest area at the municipal level, serving as a macro-scale validation metric to complement micro-scale verification based on individual sample points. By comparing the forest area derived from our deep learning framework with the officially reported totals for each municipality, potential uncertainties associated with single-source validation—such as limitations in sample representativeness—can be effectively mitigated. This multi-scale and multi-source validation strategy enhances the overall credibility of the model’s performance in forest type identification and provides a more robust assessment framework.

During the validation process, the model-inverted forest area for each municipal division across different years was quantitatively compared with official statistics. To evaluate the accuracy of regional-scale forest area estimation, the Relative Error (RE) was used as a statistical measure to quantify the misestimation ratio. This approach effectively compensates for the shortcomings of relying solely on sample-based validation and ultimately establishes an integrated, multidimensional verification system that combines point-scale (sample-based) and area-scale (statistic-based) assessments. The implementation of this comprehensive validation framework not only strengthens the reliability of our research findings but also provides a replicable methodology for similar studies in other regions.

3. Results

3.1. Qualitative Comparison with Well-Established Map Products

3.1.1. Comparing with Four Land-Cover Products

To qualitatively evaluate the forest mapping product developed in this study, we conducted a comprehensive visual comparison with four widely used land-cover products across seven representative regions in Hunan Province, covering diverse landscapes and different land-cover patterns (Figure 6). Our large-scale comparative analysis demonstrates distinct advantages of our product in capturing fine-scale forest distribution features.

In Hanshou, Huarong, and Jiangyong Counties, mountain forests exhibit complex distribution patterns characterized by terrain-adapted extensions, intricate edges, and fine fragmentation. While the original Landsat imagery (first column) clearly reveals these details, comparative products show significant limitations. For instance, GlobaLand30 (Figure 6d) presents over-smoothed forest boundaries, flattening originally tortuous and fragmented edges and resulting in substantial loss of spatial detail. Similarly, FROM_GLC2015 (Figure 6e) displays fragmented and discontinuous forest patches, failing to represent the actual continuous distribution of mountain forests. Other products (Figure 6b,c) exhibit insufficient capability in identifying forest-covered land types. In contrast, our product (Figure 6a) accurately captures the irregular, terrain-following extension of mountain forests, with natural and detailed transitions between forest edges and non-forest areas that faithfully reflect microtopographic variations.

Detailed comparison in Hanshou and Yuanjiang Counties confirms our product’s superior representation of complex spatial patterns. Landsat imagery shows forests linearly distributed along valley sides, interspersed with croplands and settlements. The north side exhibits continuous strip forests, while the south contains three small forest patches embedded in farmland, with boundaries closely following terrain. Although FROM_GLC2015 (Figure 6e) captures the valley’s main forest distribution, it shows significant spatial expansion deviations. For example, the actual wedge-shaped forest patches on the south side are erroneously simplified into continuous forests, poorly matching the valley topography. Other products (Figure 6b–d) exhibit varying classification errors, particularly misclassifying forests as non-forest categories. By comparison, our product shows significantly higher consistency with Landsat imagery, not only accurately identifying the strip forests along valley sides and small patches on the south side, but also precisely matching forest boundaries with topographic characteristics. The extension direction of strip forests on the north side aligns with valley trends, while the shape, area, and spatial relationships of eastern forest patches with surrounding non-forest features highly correspond to actual patterns, effectively restoring the spatial distribution logic of forest-non-forest interfaces in this region.

To clarify the variability observed in Figure 6, we supplemented spring (April–June) and summer (July–September) remote sensing imagery of the study area (Figure 7). Visual comparison of Figure 7 reveals pronounced seasonal differences in forest attributes linked to the patterns in Figure 6. In summer, vegetation (e.g., Cinnamomum camphora in southern Hunan) shows elevated near-infrared (NIR) reflectance, enhancing spectral discrimination from non-forest land such as croplands in the northern plain. Conversely, spring conditions induce incomplete canopy closure, increasing spectral overlap between forests and early-growing crops. In terms of spatial patterns, summer imagery depicts continuous forest boundaries consistent with peak growing-season canopy density, whereas spring imagery shows fragmented boundaries for deciduous forest patches.

3.1.2. Comparing with Three Forest Products

To qualitatively assess the forest mapping product generated in this study, we performed a systematic visual comparison against three widely adopted forest products across five representative regions, spanning diverse landscapes and heterogeneous land-cover configurations, as shown in (Figure 8). This comparative analysis facilitates evaluation of how distinct feature representation paradigms shape forest identification performance: the proposed framework emphasizes fine-grained land cover texture extraction (consistent with the spatial details in Landsat imagery)—a design tailored to address the limitations of coarse feature modeling inherent in existing reference products. Visual scrutiny of forest cover (represented by green regions) across the study areas reveals discernible performance disparities: In Region 1, the forest distribution derived from the proposed approach exhibits greater continuity and sharper boundary definition, whereas forest cover delineated by reference products appears markedly fragmented. For Region 4, the proposed approach enables more precise demarcation of forest-nonvegetation interfaces (e.g., riparian zones), while alternative products are prone to either over- or under-estimation of forest extent. Minimal inter-product performance discrepancies are observed in vegetation-homogeneous regions (e.g., Region 5); conversely, the proposed framework demonstrates enhanced adaptability in complex terrains (e.g., Region 2 and 3) characterized by heterogeneous land cover and topographic variation.

This improved identification fidelity can bolster the reliability of downstream applications, including forest carbon stock quantification and vegetation cover dynamics analysis. Subsequent investigations will further validate the framework’s performance via quantitative metrics (e.g., overall accuracy, F1-score).

Overall, our forest mapping product demonstrates three core advantages: (1) Multi-source data fusion enables higher boundary delineation precision and effective capture of fine-scale distribution characteristics, mitigating issues of boundary ambiguity. (2) Classification models constructed with diverse, reliable training samples facilitate more comprehensive and accurate representation of forest spatial distribution, reducing omissions of small patches and misclassification errors. (3) Utilization of multi-temporal remote sensing information maintains higher consistency with actual forest distribution across different temporal dimensions, providing more reliable support for dynamic monitoring of forest resources. (4) It presents favorable adaptability in regions with complex terrain-vegetation conditions (e.g., fragmented landscapes), where the continuity of mapped forest cover shows relatively more stable performance. (5) It achieves clearer differentiation between forested areas and non-vegetated regions (e.g., riparian zones), alleviating the issues of over- and under-identification, and thus enhancing applicability in scenarios with heterogeneous land cover.

3.2. Pixel-Level Evaluation Based Manual Annotation Point-Set

To ensure the reliability of the validation results, a visual interpretation process was performed for the 9000 randomly selected validation points. The spatial distribution of these points across the study area is presented in the left panel of Figure 9: the dense sampling coverage ensures representativeness across different geographical zones (e.g., mountainous, plain, and basin regions) of the study area. For a transparent demonstration of the interpretation criteria, representative samples of forest (labeled “1”) and non-forest (labeled “0”) land cover types are displayed in the right panel of Figure 9. These samples were extracted from 30 m resolution Google Earth imagery (covering the entire study area) and correspond to typical scenarios linked to topographic/climatic conditions: Non-forest samples (labeled “0”): Include areas such as croplands (distributed in low-altitude plain regions with temperate climate), built-up lands (concentrated in urban agglomerations), and bare lands (distributed in high-altitude mountainous areas with arid climate). Forest samples (labeled “1”): Cover natural forests (in humid mountainous regions with sufficient precipitation) and planted forests (in hilly areas with moderate temperature and rainfall).

This interpretation process strictly followed the land cover classification system, and the selected examples reflect the diversity of the validation samples, thereby supporting the credibility of the subsequent accuracy assessment.

Based on the validation sample set for Hunan Province introduced in Section 2.4.2, we quantitatively validated the model accuracy through visual interpretation of nearly 9000 sample points. The OA at three-year intervals from 1999 to 2023 is shown in (Figure 10), revealing distinct spatial patterns across prefectures that correlate with landscape complexity and forest management practices.

Regions with lower OA, including Huaihua, Zhangjiajie, Shaoyang, and Changde, typically feature highly heterogeneous landscapes with complex mosaics of mountainous forests, croplands, and water bodies. This spectral complexity is amplified by heterogeneous stand structures from varied management practices, such as mixed-species forests and mosaics of different age classes. In contrast, prefectures like Changsha and Hengyang achieved higher accuracy, benefiting from extensive tracts of large-scale farmland and uniformly managed plantations. The highest accuracy was in Yueyang, Zhuzhou, Hengyang, Loudi, and Changsha, where regular landscapes—such as urban green spaces, extensive farmlands, and commercial forests—under intensive management result in clear boundaries and high spectral consistency.

As shown in (Figure 11), the model performance exhibited clear temporal trends from 1999 to 2023. The OA showed a consistent upward trend, reflecting a gradual enhancement in distinguishing forest from non-forest areas. Recall remained relatively stable above 0.9 in most years, indicating a robust ability to identify actual forest areas. However, precision experienced significant fluctuations: It started at approximately 0.6 in 1999, decreased to around 0.5 in 2002, and gradually recovered to about 0.8 after 2017. The initial decline was attributable to challenging interference factors, such as spectrally ambiguous transition zones and mixed artificial-natural forests, while the subsequent recovery aligns with more standardized forest management and increasingly distinct forest features. The F1-score consistently integrated the trends of both recall and precision, while the steadily increasing kappa coefficient indicated continuous improvement in classification consistency with actual land cover.

The classification results presented in (

Table 4. Forest coverage characteristics and model evaluation metrics across different geographical regions and municipal/prefectural areas in Hunan Province.

Geographical Region	Municipal and Prefectural Region	Number of Samples	Proportion of Forest Coverage Area of Hunan	OA (%)	Recall (%)	F1-Score (%)	Kappa (%)
East	changsha	45	3.78	95.45	98.96	84.85	73.06
	Zhuzhou	58	6.84	94.25	95.12	86.5	82.79
	Xiangtan	29	1.82	91.95	76.98	63.45	46.2
West	Xiangxi	89	8.44	92.76	89.56	74.82	62.61
	Huaihua	143	15.5	91.3	91.89	77.49	64.12
	Zhangjiajie	46	4.37	91.3	94.05	69.3	58.08
Central	Loudi	40	3.35	95.28	94.23	80.89	78.34
	Shaoyang	97	13.3	91.18	87.28	75.6	70.61
South	Hengyang	67	3.57	96.02	98.21	71.96	54.73
	Yongzhou	93	7.8	93.16	93.84	74.14	70.21
	Chenzhou	77	7.85	81.67	91.96	79.63	74.36
North	Yueyang	63	6.65	94	94.71	79.74	76.06
	Changde	103	8.43	92.34	87.47	68.03	56.43
	Yiyang	50	9.25	92.86	92.75	80.63	76.56

) further highlight regional disparities across Hunan. A clear performance gap exists between prefectures, with eastern Changsha achieving the highest OA (95.45%) and southern Chenzhou the lowest (81.67%). The lower accuracy in Chenzhou stems from its challenging terrain at the northern foot of the Nanling Mountains, where pronounced topographic relief creates shadow effects that distort spectral signals. Compounding this issue, the spectral profile of the dominant Phoebe zhennan tree overlaps significantly with other land cover types, impeding clear class separation. In Western Hunan, Huaihua (143 samples) attained an OA of 91.30% and a Recall of 91.89%, while Xiangxi (89 samples) showed a similar OA of 92.76% despite a marginally lower Recall (89.56%). This comprehensive analysis confirms that accuracy variation is primarily driven by landscape complexity and management heterogeneity, where greater complexity generally predicts lower performance.

3.3. Statistical Assessment Based on Official Government Survey

Based on the statistical validation dataset described in Section 2.4.3, official forest resource survey data from 14 prefecture-level administrative regions were collected to evaluate the statistical-level performance of our forest type classification. A comparison between the forestland area derived from our study and the official statistics from the Hunan Provincial Land Greening Status Bulletins across all prefectures is provided in Figure 12. The values in the figure represent the overestimation (positive) and underestimation (negative) of forestland area by our method.

Overall, our estimates are generally consistent with the official bulletin data for most years. Discrepancies in forestland area due to classification errors exhibited distinct spatial and temporal patterns. In eastern Hunan, overestimation primarily occurred in 1999, 2002, and 2020. The overestimation in 2002, which was particularly pronounced in Changsha and Zhuzhou, resulted mainly from a discrepancy in the definition of “forestland” between the classification scheme used in this study and the official government criteria. Specifically, our study included some open woodlands and shrublands that were excluded from the forestland category in the official survey, leading to the observed positive bias.

In western Hunan, forest cover was underestimated in 1999 (mainly in Xiangxi and Huaihua) and in 2023 (concentrated in Zhangjiajie and Xiangxi), likely due to human disturbance and land changes from afforestation and development. Southern Hunan, however, showed high agreement with government data across all years. Its warm, humid climate supports stable forests—primarily natural broad-leaved and Chinese fir—that sustain high year-round vegetation cover with minimal seasonal variation. The distinct spectral signatures of these forests reduce confusion with farmland or construction land, minimizing misclassification and aligning closely with official inventory results. In contrast, northern and central Hunan experienced misestimation in some years, largely due to forest-type misclassification.

3.4. Regulatory Effects of Environmental Factor Coupling Mechanisms on Dominant Tree Species Distribution

Precipitation, temperature, and DEM do not act independently on forest growth. Instead, they collectively shape the distribution patterns of regional dominant tree species through coupled processes that regulate plant physiological metabolism, material–energy exchange, and habitat conditions (as exemplified by the Hunan Province case presented in Figure 5.

3.4.1. Water-Nutrient Regulatory Mechanism of Precipitation

As the core input component of the forest water cycle, precipitation screens tree species via the physical process of “physiological water supply-soil nutrient transport”. Sufficient precipitation maintains turgor pressure in plant mesophyll cells and stomatal apertures, ensuring high photosynthetic efficiency and rapid dry matter accumulation. Moderate precipitation dissolves soil nutrients through leaching processes while avoiding the occurrence of waterlogging stress.

The distribution patterns of Koelreuteria paniculata in Xiangtan and Camellia oleifera in Loudi verify the regulatory effect of precipitation “moderateness”. The annual precipitation of 1200–1400 mm in Xiangtan not only meets the photosynthetic water demand of K. paniculata but also shapes the slightly acidic soil environment preferred by this species. The annual precipitation of 1100–1300 mm in Loudi balances the water demand for growth of C. oleifera and the soil aeration required for root respiration, thereby avoiding hypoxic stress. This mechanism reveals the adaptive characteristics of tree species in humid regions to the “magnitude-rhythm” of precipitation.

3.4.2. Energy-Phenology Driving Mechanism of Temperature

Temperature determines plant photosynthetic/respiratory efficiency and phenological rhythms by regulating enzyme activity and molecular motion rates. Within the optimal temperature range for enzymatic reactions, the plant carbon fixation rate reaches its peak; the accumulated temperature threshold defines the growth cycle of tree species. The evergreen characteristic of Cinnamomum camphora in Changsha stems from the mean annual temperature of 17–18 °C (falling within the optimal range of photosynthetic enzymes) and accumulated temperature of 5500–5800 °C (meeting the requirements for completing the growth cycle) in its distribution area. The fruit development of C. oleifera in Loudi relies on a mean annual temperature of 16–17 °C to maintain enzymatic catalytic efficiency while avoiding heat damage caused by high temperatures. This reflects the shaping effect of air temperature on the functional traits of tree species (e.g., evergreen/deciduous attributes, fruiting capacity).

3.4.3. Hydrothermal Spatial Reconstruction Mechanism of DEM

Through the physical differentiation effects of elevation, slope gradient, and aspect, DEM reconstructs local hydrothermal conditions: elevation drives the vertical zonality differentiation of hydrothermal conditions; slope gradient regulates soil stability; aspect shapes small-scale hydrothermal differences.

The cold tolerance of Pinus massoniana in western Hunan matches the cool habitat of high elevations, and the steep slope terrain effectively prevents waterlogging. The undulating terrain in Zhuzhou not only maintains the suitable temperature for the growth of Acer palmatum but also improves soil aeration. This indicates that topography is the core driving factor of “small scale heterogeneity” in regional tree species distribution.

In summary, the distribution patterns of dominant tree species illustrated in Figure 5 are the outcomes of the coupling of multiple physical mechanisms of environmental factors. The tree species cases in specific regions represent the concretization of macroscopic mechanisms at the local scale, providing dual “mechanism-case” support for the introduction of environmental factors into regional forest mapping.

3.5. Environmental Drivers of Classification Accuracy Heterogeneity

The spatial heterogeneity of forest classification accuracy across Hunan Province (Figure 10) is not a random pattern but a direct consequence of the coupled effects of environmental factors (precipitation, temperature, DEM) on forest habitat conditions—this finding aligns with the mechanistic framework we established earlier.

As illustrated in Figure 10, cities with low OA are concentrated in regions characterized by high topographic complexity and abundant annual precipitation. Here, the steep slopes accelerate precipitation runoff, while high elevation drives a 2–3 °C reduction in mean temperature (relative to lowland areas) and increases air humidity. This topography-mediated redistribution of hydrothermal conditions fragments forest habitats: alpine coniferous forests (e.g., P. massoniana) coexist with cold-tolerant shrubs in micro-topographic niches, creating a mosaic of vegetation types with overlapping spectral signatures. For example, the reflectance of shaded P. massoniana needles (in steep north-facing slopes) can be confused with that of dense shrubs, increasing classification uncertainty. This shows that topographic complexity amplifies spectral confusion in mountainous forest mapping.

In contrast, cities with high OA are located in low-relief plains characterized by moderate precipitation (1300–1500 mm) and stable mean temperatures (17–18 °C). The flat terrain reduces runoff and facilitates uniform infiltration of precipitation into deep, well-drained soils; meanwhile, the optimal temperature range for C. camphora’s photosynthetic enzymes fosters the development of continuous, monodominant evergreen broad-leaved forests. This homogeneous hydrothermal setting minimizes spectral variability: C. camphora canopies show consistent reflectance (low intraspecific spectral variation) and distinct signatures from non-forest types (e.g., croplands), thereby reducing classification errors. This confirms that stable habitats driven by low topographic heterogeneity directly improve forest classification accuracy.

Notably, intermediate OA values correspond to regions with moderate environmental complexity. Here, the undulating terrain creates mild hydrothermal gradients (rather than extreme fragmentation), supporting relatively continuous forest communities (e.g., K. paniculata dominant) with limited spectral overlap—this “gradient effect” of environmental factors on accuracy fills a gap in existing studies, which often focus on binary (high/low) complexity rather than continuous variation.

In summary, the accuracy pattern in Figure 10 reveals a clear causal chain: environmental factors → habitat complexity → spectral variability → classification accuracy. This mechanistic link not only explains the spatial heterogeneity of our results but also proposes a generalizable framework for forest mapping in heterogeneous regions. Incorporating high-resolution environmental variables (e.g., microtopographic DEM derivatives) can alleviate spectral confusion in complex habitats, while niche-stratified training samples may further enhance classification accuracy in mountainous areas.

4. Discussion

Our study demonstrates the practical feasibility of deriving time-series forest distributions in Hunan Province using a deep-learning approach that integrates 25-year multispectral imagery and environmental features at 30 m resolution. The model was rigorously validated through (1) qualitative comparison with four established land-cover products (CLCD, GlobaLand30, FROM_GLC2015, GLC_FCS30), showing improved boundary precision and detail preservation; (2) cross-validation with official statistical data; and (3) manual verification using 9000 random points, achieving over 92% OA across all 14 prefectural cities. These methodological strengths enable a reliable capture of forest dynamics from 1999 to 2023.

4.1. Forest Distribution in Hunan Province

To our knowledge, this study represents the first implementation of a deep learning framework integrated with long-term time-series data for forest classification in China’s Hunan Province. Previous investigations into the spatial patterns of forests in this region have predominantly relied on conventional machine learning methods, particularly the RF algorithm [66]. While effective, RF and similar approaches are highly dependent on the quality and balance of the training data and have limited capacity to resolve complex, non-linear spectral features, often leading to suboptimal accuracy, especially in spectrally heterogeneous landscapes.

In this study, we efficiently processed a total of 4257 remote sensing images using a deep learning model. As illustrated in (Figure 1c), the derived forest distribution reveals a distinct spatial pattern for Hunan Province: greater coverage in the west than in the east, and higher density in the south than in the north. This pattern is shaped by a combination of the province’s horseshoe-shaped topography and anthropogenic activities. Steep slopes in the western and southern mountains have favored forest conservation and regrowth, while the northern plains are dominated by intensive agriculture. Since around 2000, national ecological initiatives like the “Grain for Green Program” and the “Natural Forest Protection Program” have been the primary drivers of consistent forest expansion. A notable increase of 17,034.26 km² in forest cover occurred in the Hilly Region of Central Hunan between 1999 and 2023, largely due to the expansion of planted forests.

Our U-Net-based forest distribution product demonstrates superior consistency with actual forest patterns when qualitatively compared against four established land-cover products (CLCD, GlobaLand30, FROM_GLC2015, GLC_FCS30). This improvement can be attributed to the fundamental methodological difference. The reference products are designed for large-scale, multi-category land cover mapping—a far more complex task—and primarily rely on traditional machine learning, which has more limited feature-extraction capabilities than deep learning. The strong performance of our model is further evidenced by its high cross-temporal adaptability. Although trained on samples collected in 2023, it successfully generated accurate forest data for three-year intervals from 1999 to 2023. This was rigorously validated through manual verification of 9000 random points, which indicated high spatial consistency, and a steady increase in the Kappa coefficient from 1999 to 2023 based on field-survey-derived samples.

The steady growth and optimization of Hunan’s forest ecosystems, as captured by our long-term analysis, underscore the success of decades of ecological protection. These efforts have not only enhanced the natural environment but also bolstered regional sustainable development. For instance, economic forests have significantly improved household incomes, and the province’s 12.73 million hectares of forestland now constitute a critical carbon sink for southern China, playing a vital role in achieving China’s “dual carbon” goals (carbon peaking and carbon neutrality).

4.2. Seasonal Phenological Effects on Forest Classification Variability

The spatial variability in forest classification results is closely related to seasonal effects, as supported by the spring-summer imagery comparison (Figure 7). Evergreen species (e.g., Cinnamomum camphora) maintain stable spectral signals across seasons, while deciduous species (e.g., Ginkgo biloba) exhibit significant spectral shifts from spring to summer—this inconsistency contributes to heterogeneous classification outcomes.

Regions with higher classification precision (e.g., southern mountainous areas) align with summer-dominated data, where strong spectral contrast between forests and non-forest land enhances classification accuracy. In contrast, variable classification performance in regions like northern agricultural zones corresponds to spring-dominated data, where incomplete canopy closure increases spectral overlap between forests and early-growing crops. This confirms that seasonal phenology is a key factor influencing the stability of forest classification results.

4.3. Performance Benchmarking Against Random Rorest and NDVI Thresholding Methods

To further validate the performance of the proposed deep learning framework in forest mapping, we conducted a visual comparison with two traditional methods (RF and NDVI thresholding) using the same Landsat data inputs, as presented in Figure 13. This figure displays results from two representative regions (two rows) across six columns: Landsat reference imagery, our deep learning results, RF outputs, and NDVI thresholding results with thresholds of 0.3, 0.4, and 0.5 (green areas represent forest cover).

The visualization reveals distinct performance disparities. First, the spatial continuity and boundary accuracy. The deep learning results (second column) exhibit highly continuous forest cover that aligns closely with the texture and structure of the Landsat reference imagery (first column). In contrast, the RF outputs (third column) appear fragmented, with blurred boundaries and disconnected forest patches—this is likely due to FR’s reliance on handcrafted features, which fail to capture the fine spatial patterns of heterogeneous landscapes (e.g., the mountainous terrain in the first row). Second, the sensitivity of NDVI thresholding to parameter selection. The NDVI thresholding results (columns 4–6) show clear dependence on threshold values: a low threshold (0.3) leads to over-identification (excessive green areas, including non-forest regions); a high threshold (0.5) causes under-identification (sparse green areas, omitting large portions of actual forest); even the moderate threshold (0.4) fails to match the spatial consistency of the deep learning results.

This instability makes NDVI thresholding unsuitable for scenarios where environmental conditions (and thus NDVI values) vary across time or space. Third, the implications for long-term sequential mapping. While the RF and NDVI thresholding methods have lower computational costs, their limitations (fragmentation, threshold sensitivity) directly hinder their application in long-term forest mapping. In contrast, the deep learning framework’s ability to maintain continuous, boundary-accurate results (consistent with reference imagery) is critical for capturing temporal changes in forest cover—this advantage aligns with our study’s core goal of multi-decadal sequential mapping.

4.4. Driving Mechanisms of Forest Resource Dynamics in Hunan Province (1999–2023)

As a typical subtropical region with high carbon sequestration capacity in China, the dynamic evolution of forest resources in Hunan Province bears both regional particularities and national strategic significance. The recent trends in forest changes and driving mechanisms are highly aligned with the long-term study period (1999–2023) adopted in this research. In terms of change trends, Hunan Province has witnessed steady and continuous growth in forest area since 1999, with the forest coverage rate increasing from 52.7% in 2000 to 59.9% in 2023. Concurrently, the quality of the forest ecosystem has been optimized in tandem, and the volume per unit area of arbor forests has risen remarkably.

The core driving factors underlying these changes can be categorized into two dimensions: first, policy-driven initiatives. A series of national ecological restoration projects, such as the Grain for Green Program and the Natural Forest Protection Program, launched around 2000, have been fully implemented across Hunan Province. The cumulative area of converted farmland to forest has exceeded 1 million hectares, effectively curbing the tendency of deforestation for arable land and facilitating the natural restoration and artificial reconstruction of forest ecosystems. Second, synergistic effects of anthropogenic activities and economic development. On the one hand, urbanization and intensive agricultural development have encroached on forest areas in some local regions; on the other hand, the optimization of the forestry industrial structure (e.g., economic forest cultivation) and the enhancement of ecological protection awareness have further consolidated the outcomes of forest growth.

The selection of the 1999–2023 period in this study precisely covers the full implementation cycles of the aforementioned policies and the critical stages of forest dynamic evolution. This temporal scope enables the systematic capture of forest change processes driven by the interaction of policy interventions and anthropogenic activities, thereby providing accurate temporal-scale support for evaluating the effectiveness of regional ecological projects.

4.5. Strengthening Forest Model Performance Evaluation Based on the Coupling of Field Observations and Remote Sensing

This integrated approach significantly enhances the reliability of forest resource monitoring across heterogeneous landscapes, thereby providing a more scientific and practical basis for the implementation of forest inventory, forestry planning, and ecological conservation projects. Consequently, this methodological framework offers valuable insights for relevant authorities to formulate targeted management strategies, facilitating the balance between ecological integrity and socioeconomic development needs.

On the one hand, field survey data serve as a high-precision ground truth benchmark for sample points, covering multiple land cover types (e.g., forestland, cropland, and construction land) as well as key forest community attributes, including species composition, stand structure, canopy density, and growth status. This fundamentally ensures the accuracy and representativeness of random point annotation across diverse ecological zones. On the other hand, by coupling precise spatial location information and detailed field-derived attributes with multi-temporal spectral and textural features extracted from remote sensing imagery, field survey data support the construction of a comprehensive validation system, which effectively bridges the scale gap between plot-based measurements and pixel-level classifications.

With the application of this integrated framework, quantitative evaluation metrics—such as OA and recall rate—exhibit not only sound statistical reliability but also the capacity to accurately capture the spatiotemporal heterogeneity inherent in real-world landscapes. This, in turn, strengthens the validity of model performance assessment under varying ecological conditions.

4.6. Limitations and Future Plans

This study has several limitations that point to valuable directions for future research. First, our methodology was confined to the U-Net architecture for deep learning-based forest extraction. Other advanced deep learning models, as well as hybrid approaches that integrate deep learning with traditional machine learning, were not explored and could potentially enhance classification accuracy. Second, the analysis relied exclusively on Landsat imagery at a 30 m spatial resolution. While Landsat is indispensable for long-term time-series analysis (e.g., from 1985 onward), future work will focus on the spatiotemporal fusion of Landsat with higher-resolution Sentinel-2 (10 m) imagery. This is expected to advance long-term forest mapping to a finer spatial granularity. Furthermore, the integration of multi-resolution remote sensing data and the inclusion of additional feature sets into deep learning models present promising avenues for improving tree species identification and fine-grained forest classification.

Future efforts will therefore focus on the following:

(1)

The prevailing approach to deep learning-based forest extraction has been largely centered around the U-Net architecture. A recognized challenge with this framework is its constrained capacity for modeling long-range contextual relationships, such as continuous ecological boundaries, and for effectively fusing diverse, multi-source features. To address this, future work will focus on constructing hybrid deep learning frameworks that integrate U-Net with Vision Transformer (ViT) and traditional machine learning features. Framework Design, Embed the local feature extraction capability of U-Net (via encoder-decoder skip connections) with the global context modeling of ViT (via multi-head self-attention), and further incorporate hand-crafted features (e.g., GLCM texture features, NDVI time-series trends, and topographic attributes from DEM) as auxiliary input channels. Comparative Validation, Conduct systematic comparisons among the proposed hybrid framework, baseline models (U-Net, SegFormer, ResUNet), and traditional machine learning methods (RF, SVM) across multiple independent datasets—covering diverse forest types (coniferous, broadleaf, mixed forests) and complex terrains (mountainous, plain, and riparian zones). Evaluation Metrics: Assess performance using metrics such as overall accuracy (OA), weighted F1-score, Kappa coefficient, and intersection over union (IoU) for forest patches; additionally, quantify model robustness in small-sample scenarios (e.g., sparse forest areas). This work is expected to improve the IoU of forest extraction by 5–8% and enhance the generalization ability of the model across heterogeneous landscapes.

(2)

The current analysis is constrained to 30 m Landsat imagery, which lacks sufficient spatial detail for fine-grained forest mapping (e.g., small forest patches or edge zones). Future research will focus on spatiotemporal fusion and synergistic utilization of multi-source remote sensing data: Employ advanced spatiotemporal fusion algorithms to integrate long-term Landsat time-series (1985–present) with 10-m Sentinel-2 optical imagery, generating continuous, high-spatiotemporal-resolution (10 m, 16-day) surface reflectance datasets. Meanwhile, incorporate Sentinel-1 C-band SAR data (VV/VH polarization features) to compensate for optical data gaps caused by clouds/rain (e.g., rainy seasons in the Dongting Lake basin), and fuse SRTM DEM-derived topographic features (slope, aspect) to distinguish terrain-driven forest type variations (e.g., sun-facing vs. shade-facing slope vegetation). Evaluate the fused data in scenarios including fine-scale forest boundary extraction, sub-compartment-level forest classification, and disturbance detection (e.g., small-scale deforestation). Compare the performance of different fusion strategies (single-temporal vs. time-series fusion) in improving classification accuracy. This effort aims to refine the spatial granularity of long-term forest mapping from 30 m to 10 m, while enhancing classification stability in cloud-prone or topographically complex regions.

(3)

The current framework primarily addresses forest/non-forest extraction and broad forest type mapping, yet accurate discrimination of fine-grained stand types—such as evergreen broadleaf, deciduous broadleaf, coniferous forests, and their closed-canopy subtypes—remains a significant challenge. This limitation stems from the high spectral similarity among tree species, seasonal phenological variations, and complex canopy structures. Future research will prioritize developing a dedicated classification branch within the hybrid deep learning framework to address this. The methodology will involve: Feature Enrichment, integrating multi-temporal spectral indices, textural features from VHR imagery (e.g., GLCM from fused 10-m data), and vertical structure information (where available, from GEDI or terrain-corrected metrics). Hierarchical Classification Strategy, implementing a cascaded model that first separates forest/non-forest, then discriminates between major life forms (coniferous vs. broadleaf), and finally classifies subordinate stand types (evergreen/deciduous, closed/open) using targeted feature sets and potentially multi-task learning. Physically Guided Modeling, incorporating species distribution constraints based on bioclimatic variables (temperature, precipitation) and topographic factors (elevation, slope, aspect) as prior knowledge or auxiliary inputs to refine ecologically implausible predictions. Validation will be conducted using carefully compiled ground-truth datasets from forest inventories and field surveys, with performance assessed via class-specific precision, recall, F1-score, and overall accuracy. This targeted effort aims to achieve a stand-type classification accuracy exceeding 85% for major classes, providing a more ecologically meaningful product for biodiversity assessment, carbon stock modeling, and precision forestry management.

5. Conclusions

Based on a 25-year (1999–2023) time-series analysis of 30 m Landsat imagery integrated with auxiliary environmental data—including DEM, temperature, and precipitation—this study successfully reconstructed and analyzed the spatiotemporal dynamics of forest distribution across the entire Hunan Province, China ( 211,800 km²). By developing and applying a deep learning framework trained on meticulously annotated samples, we generated a consistent, high-accuracy historical forest cover dataset at three-year intervals. This long-term, large-scale mapping effort provides a detailed and reliable account of forest changes, capturing nuanced patterns that are often missed by coarser global products.

The core methodological advancement of this work lies in its effective multisource data fusion strategy. Moving beyond reliance on optical spectral information alone, our model synergistically incorporates topographic and climatic variables. This approach enhances the disambiguation of spectrally similar land-cover types (e.g., certain agricultural lands versus sparse forests) and better represents forest growth constraints imposed by terrain and climate, leading to more physiographically plausible results. Qualitative comparisons with several existing regional and global forest products (e.g., FROM-GLC, Hansen Global Forest Change) across seven representative sub-regions consistently demonstrated the superior capability of our product in resolving fine-scale spatial patterns, such as narrow riparian forests, complex patch edges, and small-scale afforestation/deforestation patches.

Quantitative validation based on 9000 manually interpreted, statistically stratified random points provided a robust, pixel-level accuracy assessment. The model achieved high and spatially consistent performance, with average overall accuracies (OAs) of 93.9% (East), 91.8% (West), 93.2% (Central), 90.5% (South), and 90.2% (North). These results not only confirm the model’s precision but also indicate its stable generalizability across diverse geographical and ecological settings within the province. Furthermore, a trend-level consistency check revealed strong agreement between our mapped forest area trends and official forest inventory statistics over the past two decades, significantly reinforcing the credibility and practical utility of our data series for regional-scale analyses.

In contrast, cities with high OA are concentrated in low-relief plains featuring moderate precipitation (1300–1500 mm) and stable mean temperatures (17–18 °C). Flat terrain and favorable hydrothermal conditions facilitate the contiguous development of monodominant evergreen broad-leaved forests, reducing spectral variability and classification errors. This confirms that stable habitats maintained by low topographic heterogeneity directly enhance forest classification accuracy. Beyond this localized observation, the synergistic effects of precipitation, temperature, and DEM-derived topographic metrics are identified as universal drivers regulating forest classification accuracy across the entire Hunan Province: favorable hydrothermal combinations and gentle terrain mitigate spectral confusion between forest and non-forest classes, whereas extreme climatic conditions and rugged topography exacerbate such ambiguities. Comparative experiments with the RF algorithm further confirm that our proposed deep learning framework outperforms traditional machine learning methods in delineating fine-scale forest boundaries and preserving spatial continuity—an advantage particularly prominent in heterogeneous mountainous areas where RF tends to yield fragmented mapping results. Moreover, cross-validation against existing regional and global forest products demonstrates that our dataset more precisely captures small-scale forest dynamics typically overlooked by coarse-resolution products, with consistent performance superiority.

Beyond producing a validated dataset, this study elucidates key forest dynamics in a crucial ecological zone. The 25-year maps reveal the net effects of major ecological engineering projects (e.g., the Grain-for-Green Program), urban expansion, and sustainable forestry practices, offering empirical evidence for evaluating past policy impacts. In summary, this research establishes a robust scientific framework for generating reliable long-term forest cover data. The framework underscores the transformative potential of integrating multi-source remote sensing data with deep learning to overcome long-standing challenges in large-scale environmental monitoring.

The resulting dataset and methodology create a solid foundation for numerous downstream applications. These include, but are not limited to the following: (1) Refining provincial- and national-scale carbon stock and flux estimations; (2) informing spatially explicit ecological conservation planning and biodiversity habitat assessment; (3) modeling hydrological services and soil erosion risks related to forest change; and (4) providing a benchmark for validating and calibrating coarser-resolution global models. Ultimately, this work highlights the critical value of advanced remote sensing analytics in addressing complex socio-ecological systems, providing essential evidence-based insights for achieving sustainable forest management and climate change mitigation goals under mounting anthropogenic and environmental pressures.